This page contains a Flash digital edition of a book.
HPC virtualisation


Virtualisation, of course, doesn’t only


apply to processing, but to communications and storage, where it does have a role to play. ‘You might have a storage area network


that you’re writing to and you’ve no idea what the physical capacity or structure of the data store is, you’re just using it to store data. Virtualisation is quite popular there, for HPC, [as a way of] sharing information for a wide number of tasks,’ says Osborne. And in an interesting twist, Shai


Fultheim, founder and CEO of ScaleMP, has taken that fact and turned it on its head. ScaleMP has taken the way storage is virtualised and expanded it into compute. While standard compute virtualisation


isn’t suitable for HPC, Fultheim says, ‘every storage system that’s used today is virtualised. If you buy something from EMC and open it up, it’s one big box of virtualised storage.’ ScaleMP, therefore, focuses on


‘virtualisation for aggregation’; taking multiple machines and making them work together. ‘We take many machines and virtualise


them as one, allowing customers to run workloads that need hundreds of cores and many, many terabytes of memory. And in HPC this is exactly what you need. If you have an application that needs a lot of memory, [through virtualisation for aggregation] we can create a machine that has four terabytes of memory and you can go and run it. To buy a machine with that memory would cost millions of dollars.’ Likewise with CPUs, Fultheim says.


‘Applications needing 500 cores – to buy a 24 SCIENTIFIC COMPUTING WORLD


machine like that will cost you a fortune and a half. But if you take multiple machines and aggregate them to one, now you have the ability to run it.’ Another issue to consider, as in any


other virtual set up, is licensing – software vendors are still focused on a per-processor licence, which doesn’t work well in a virtualised world. ‘If you want to buy [software] for


doing computational fluid dynamics, these packages today are very expensive, and you pay for them on a number-of-processors basis, tied to a particular computer IP address,’ says Parsons, at EPCC. ‘Doing it in a virtualised environment,


that can change every time you use it, so the licensing models don’t really work. They just haven’t really caught up. And the majority of [the vendor’s] revenue comes from licensing, so they don’t want to do anything that damages their ability to make money,’ he says. Osborne agrees. ‘[Vendors] still think


there should be a licence for every piece of hardware and it’s a nonsense in this world.


CLOUD HASN’T TAKEN OFF IN


HPC TO ANY GREAT EXTENT, FOR SEVERAL REASONS… BUT THAT’S LIKELY TO CHANGE OVER TIME – CLOUD VENDORS ARE ALREADY RECOGNISING THE HPC MARKETPLACE


It’s going to be impossible to find out at any given time where an instance occurs, and it may only be there fleetingly anyway.’ There is a move towards dynamic


licence servers, which let licences be shared, Osborne says, but it remains an issue for users of virtualisation – in all fields, not just HPC. Overall, it seems HPC data centre managers aren’t likely to shift to virtualisation as long as they can do without it – the overheads will always be too high on in-house servers. The one area that might see a shift, though, is cloud computing – by its nature a virtualised environment. Cloud hasn’t taken off in HPC to any


great extent, for several reasons – the built-in overheads affecting speed, latency issues, the problems in moving large data sets, and basic issues of not really knowing what’s happening with your data. But that’s likely to change over time – cloud vendors are already recognising the HPC marketplace and working to provide services that fit its needs. Inevitably there will be some HPC users who will choose that route and take their processing to the cloud, balancing the downsides against the scalability and simplicity of using the cloud. ‘All these things have their place,’


says Osborne. ‘It’s a “horses for courses challenge”. Virtualisation is hugely attractive in things like test development, and anywhere performance is not at a premium. But if you’re trying to do something deterministic or produce a simulated output or something in real time then you really can’t afford to do it that way.’


www.scientific-computing.com


Page 1  |  Page 2  |  Page 3  |  Page 4  |  Page 5  |  Page 6  |  Page 7  |  Page 8  |  Page 9  |  Page 10  |  Page 11  |  Page 12  |  Page 13  |  Page 14  |  Page 15  |  Page 16  |  Page 17  |  Page 18  |  Page 19  |  Page 20  |  Page 21  |  Page 22  |  Page 23  |  Page 24  |  Page 25  |  Page 26  |  Page 27  |  Page 28  |  Page 29  |  Page 30  |  Page 31  |  Page 32  |  Page 33  |  Page 34  |  Page 35  |  Page 36  |  Page 37  |  Page 38  |  Page 39  |  Page 40  |  Page 41  |  Page 42  |  Page 43  |  Page 44  |  Page 45  |  Page 46  |  Page 47  |  Page 48  |  Page 49  |  Page 50  |  Page 51  |  Page 52